Data-driven surrogate models for forecasting experimentally measured fluid flows

arXiv:2606.10848 · physics.flu-dyn · Submitted 2026-06-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Data-driven surrogate models for forecasting experimentally measured fluid flows".

Mira: Here's the gist of the paper. Data-driven surrogate models can provide meaningful predictions over short time horizons and propagate low-frequency dynamics over longer forecast periods while struggling to preserve transient flow features and high-frequency energy content when faced with noisy measurements and incomplete state observations Modeling Approaches Evaluated The study analyzed four common data-driven model architectures to represent a range of modalities surrogate modeling performance, including fully convolutional neural networks (FCNs), U-Nets, Fourier neural operators (FNOs), and dynamic mode decomposition (DMD) The models were trained on velocity fields measured in the near wake of a circular cylinder at a range of Reynolds numbers U-Nets and FCNs were included to represent common, generalized deep learning frameworks Fourier neural operators (FNOs) were included as one of the most common architectures in the growing field of operator learning Dynamic mode decomposition (DMD), which is not a deep learning method, has been a common modeling tool in the fluid mechanics community for decades Data Acquisition and Preprocessing The data was collected using a free-surface water tunnel with test section dimensions of 0.15 m (W) × 0.15 m (H) × 0.61 m (L) An acrylic cylinder of diameter D = 9.53 × 10−3 m was mounted to span the walls of the test section approximately equidistant from the free surface and the tunnel floor High-speed video was recorded via an IDT XSM-3520 high-speed camera, with resolution 2144 px × 1440 px The non-dimensional formation time, given as t∗ = U t/D, was used to normalize the flow development between subsequent velocity fields across Reynolds numbers The resulting velocity fields have grid spacing of 9.388 × 10−4 m (< 0.1D) after processing with a 32 × 32 pixel interrogation window with a 50% overlap Performance Across Forecast Horizons When examining performance over extended forecast horizons at a fixed Reynolds number (Re = 590), the predictions at the first timestep generally show low error and a high level of spatial detail By the tenth time-step, the forecast velocity fields appear significantly less detailed than the actual field, and more significant error has accumulated The error accumulates most rapidly over the first several prediction time-steps for the U-Net, FNO, and DMD models However, repeated recursive application of the surrogate model does not accelerate error accumulation but rather stabilizes it at longer time horizons Impact of Flow Physics and Reynolds Number The performance is most sensitive to Reynolds number from Re = 230 to Re = 830 for all four models As the Reynolds number increases, turbulent fluctuations become more energetic and three-dimensional Irregular out-of-plane velocity fluctuations at large Reynolds numbers are especially problematic because they cannot be measured with the planar two-component PIV The effective result is more relative smoothing in the predicted velocity fields, which effectively truncates high-frequency dynamics making the resulting error less sensitive to Reynolds number Error Decomposition and Physical Insight Through a triple decomposition, it was established that the majority of the error comes from the high-frequency fluctuating velocity component The error is dominated by that of the irregular velocity fluctuations The estimated planar turbulent kinetic energy (TKE) in the predictions decays nonlinearly, with the most rapid decrease happening over the first few time-steps for U-Net and FNO surrogates The linear DMD surrogate models are able to maintain more energy in predictions across intermediate forecast horizons The data-driven models are able to propagate low-frequency dynamics over long forecast horizons tested at Re = 590 However, fine high-frequency spatial details decay rapidly over the first several prediction time-steps Given the chaotic nature of the turbulent dynamics, it seems unlikely that any surrogate would be capable of accurately forecasting the development and evolution of secondary instabilities and finer scale flow features across moderate time horizons The models show potential for faster-than-real-time evaluation, possibly enabling real-world deployment for predictive model-based flow control Author Contributions P.R.

Kai: and M.G.

Mira: conceptualized this work P.R.

Kai: and C.W.

Title and authors: Kai: Moving on to what the paper actually says about its main finding, this paper, "Data-driven surrogate models for forecasting experimentally measured fluid flows", it shows that these data-driven models show promise for making faster than real-time forecasts of fluid flows.

Mira: They are showing this promise specifically because they are dealing with real world engineering applications where you have limited, imperfect and incomplete experimental measurements. That’s the key constraint they’re working under >

Lev: And what's important is that they trained these models on a dataset of two dimensional, twocomponent particle image velocimetry measurements to forecast the time evolution of experimentally measured cylinder wakes >

Kai: So, the researchers used this dataset to train fully convolutional neural networks, U-Nets, Fourier neural operators and dynamic mode decomposition based models to predict how velocity fields develop >

Mira: The goal was to see which of these methods could best handle those transient flow features and the limited observations when forecasting over extended time horizons at a fixed Reynolds number, like Re equals five hundred ninety > <ref:2606.10848#pg2,at a fixed Reynolds number>

Lev: They examined the performance by training the models across a range of Reynolds numbers, from Re equals two hundred thirty to Re equals two thousand nine hundred twenty to see how turbulent and three dimensional flow phenomena affect their forecast accuracy > <ref:2606.10848#pg2,turbulent and three dimensional flow phenomena>

The paper's summary: Kai: The paper summarizes that they analyzed a wide array of data-driven methods for predicting the time evolution of fluid flows. They point out that models like those from Srinivasan et al. and Nakamura et al. have matched statistical quantities well, but struggle with making accurate instantaneous predictions >

Mira: They also mention Wu et al.'s proposal of a multi-resolution convolutional interaction network to predict cylinder flow data from eddy viscosity turbulence models, but they found basic linear dynamic mode decomposition was better for a cylinder flow with constant inlet velocity >

Lev: And Fukami et al. combined a convolutional neural network auto encoder with sparse identification of nonlinear dynamics, or SINDy, to model both laminar and shear flow data generated by direct numerical simulation >

Kai: The paper highlights that these existing methods are sensitive to noisy observations and often need problem specific handling when used for this kind of prediction >

Mira: That sensitivity is what motivates this work; they are specifically trying to characterize data-driven approaches that can contend with transient flow features and limited, imperfect observations >

The paper's improvements: Lev: The authors suggest a few ways to improve these models. First, they point out that the performance is most sensitive to Reynolds number from Re equals two hundred thirty to Re equals eight hundred thirty for all four models >

Kai: And they suggest a way to deal with the high frequency stuff. They say that the majority of the error comes from that high frequency fluctuating velocity component, and this error is dominated by irregular velocity fluctuations >

Mira: So, one improvement is to incorporate physics informed learning constraints into data driven losses. This would be like making sure the models follow known fluid dynamics assumptions, such as divergence free velocity fields even when the experimental data is incomplete >

Lev: Another idea comes from looking at DMD; they suggest implementing a method for debiasing Dynamic Mode Decomposition models by explicitly prioritizing the most energetic modes through singular value decomposition to get a better estimate of planar turbulent kinetic energy >

Conclusion: Kai: So to wrap up, the paper on "Data-driven surrogate models for forecasting experimentally measured fluid flows" shows that these data driven models can propagate low frequency dynamics over long forecast horizons, but fine high frequency spatial details decay very rapidly over the first few prediction time steps >

Mira: Given that turbulent dynamics are chaotic, they don't think any surrogate model will be able to accurately forecast secondary instabilities or finer scale flow features across moderate time horizons >

Lev: They do show potential for faster than real time evaluation, which means they could be used for predictive model based flow control in the real world >

Kai: And one thing is that the performance is most sensitive to Reynolds number from Re equals two hundred thirty to Re equals eight hundred thirty for all four models >

Mira: So, while we see some wins, the paper also flags that they need better ways to handle those high frequency fluctuations when making predictions >

Lev: Ultimately, the paper on "Data-driven surrogate models for forecasting experimentally measured fluid flows" is a look at what these tools can do right now and where they still need more refinement before they are fully deployed >

California Institute of Technology · Albert Nerken School of Engineering · University of Iowa

physics.flu-dyn

Submitted: 2026-06-09

Updated: 2026-11-10

Journal ref: Physics of Fluids 38, 105145 (2026)

DOI: 10.1063/5.0354356

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 86/100

The gist: Data-driven surrogate models can provide meaningful predictions over short time horizons and propagate low-frequency dynamics over longer forecast periods while struggling to preserve transient flow…

Key concepts

Surrogate Models
These are AI models trained on experimental data to quickly predict future states of a fluid flow without needing expensive, real-time physical simulations. They aim to mimic the complex fluid behavior observed in the measurements.
Reynolds Number (Re)
This dimensionless number measures the ratio of inertial forces to viscous forces in a fluid. The study found that performance is most sensitive between Re=230 and Re=830, as higher Re creates more energetic, three-dimensional turbulence that is harder for the models to capture.
High-Frequency Dynamics
These refer to the small-scale, rapidly changing features of the fluid flow velocity field. The research consistently showed that these fine details decay quickly in predictions over time, meaning surrogate models lose accuracy in capturing these rapid fluctuations.
Dynamic Mode Decomposition (DMD)
A traditional modeling tool used here that is not a deep learning method. It decomposes the flow data into different modes of motion. The study suggests debiasing DMD using Singular Value Decomposition can help it better estimate important physical energy components.

Terminology

Summary

Data-driven surrogate models can provide meaningful predictions over short time horizons and propagate low-frequency dynamics over longer forecast periods while struggling to preserve transient flow features and high-frequency energy content when faced with noisy measurements and incomplete state observations

Modeling Approaches Evaluated

The study analyzed four common data-driven model architectures to represent a range of modalities surrogate modeling performance, including fully convolutional neural networks (FCNs), U-Nets, Fourier neural operators (FNOs), and dynamic mode decomposition (DMD) The models were trained on velocity fields measured in the near wake of a circular cylinder at a range of Reynolds numbers U-Nets and FCNs were included to represent common, generalized deep learning frameworks Fourier neural operators (FNOs) were included as one of the most common architectures in the growing field of operator learning Dynamic mode decomposition (DMD), which is not a deep learning method, has been a common modeling tool in the fluid mechanics community for decades

Data Acquisition and Preprocessing

The data was collected using a free-surface water tunnel with test section dimensions of 0.15 m (W) × 0.15 m (H) × 0.61 m (L) An acrylic cylinder of diameter D = 9. p. 5

Where the paper says it

15 m (H) × 0. An acrylic cylinder of diameter D = 9.

p. 5
Where the paper says it

15 m (H) × 0. An acrylic cylinder of diameter D = 9.

53 × 10−3 m was mounted to span the walls of the test section approximately equidistant from the free surface and the tunnel floor High-speed video was recorded via an IDT XSM-3520 high-speed camera, with resolution 2144 px × 1440 px The non-dimensional formation time, given as t∗ = U t/D, was used to normalize the flow development between subsequent velocity fields across Reynolds numbers The resulting velocity fields have grid spacing of 9. p. 5
Where the paper says it

High-speed video was recorded via an IDT XSM-3520 high-speed camera, with resolution 2144 px × 1440 px. To maintain a near-constant non-dimensional inter-sample time at various tunnel speeds, the frame rate of the camera was adjusted between Reynolds numbers.

388 × 10−4 m (< 0. p. 6
Where the paper says it

fields have grid spacing of 9. 388 × 10−4 m (< 0.

p. 6
Where the paper says it

fields have grid spacing of 9. 388 × 10−4 m (< 0.

1D) after processing with a 32 × 32 pixel interrogation window with a 50% overlap

Performance Across Forecast Horizons

When examining performance over extended forecast horizons at a fixed Reynolds number (Re = 590), the predictions at the first timestep generally show low error and a high level of spatial detail By the tenth time-step, the forecast velocity fields appear significantly less detailed than the actual field, and more significant error has accumulated The error accumulates most rapidly over the first several prediction time-steps for the U-Net, FNO, and DMD models However, repeated recursive application of the surrogate model does not accelerate error accumulation but rather stabilizes it at longer time horizons

Impact of Flow Physics and Reynolds Number

The performance is most sensitive to Reynolds number from Re = 230 to Re = 830 for all four models As the Reynolds number increases, turbulent fluctuations become more energetic and three-dimensional Irregular out-of-plane velocity fluctuations at large Reynolds numbers are especially problematic because they cannot be measured with the planar two-component PIV The effective result is more relative smoothing in the predicted velocity fields, which effectively truncates high-frequency dynamics making the resulting error less sensitive to Reynolds number

Error Decomposition and Physical Insight

Through a triple decomposition, it was established that the majority of the error comes from the high-frequency fluctuating velocity component The error is dominated by that of the irregular velocity fluctuations The estimated planar turbulent kinetic energy (TKE) in the predictions decays nonlinearly, with the most rapid decrease happening over the first few time-steps for U-Net and FNO surrogates The linear DMD surrogate models are able to maintain more energy in predictions across intermediate forecast horizons

Conclusion on Model Capabilities

The data-driven models are able to propagate low-frequency dynamics over long forecast horizons tested at Re = 590 However, fine high-frequency spatial details decay rapidly over the first several prediction time-steps Given the chaotic nature of the turbulent dynamics, it seems unlikely that any surrogate would be capable of accurately forecasting the development and evolution of secondary instabilities and finer scale flow features across moderate time horizons The models show potential for faster-than-real-time evaluation, possibly enabling real-world deployment for predictive model-based flow control

Author Contributions

P.R. and M.G. conceptualized this work P. p. 1

Where the paper says it

, flow control), models must contend with limited, imperfect, and incomplete experimental measurements. In this work, we present an analysis of data-driven surrogate models trained to forecast the time-evolution of experimentally measured cylinder wakes in the subcritical vortex shedding regime.

R. and C.W. collected the data for this work P. p. 5
Where the paper says it

phenomena that is increasingly challenging to both model and observe, while maintaining a common dominant flow feature in the primary vortex shedding periodicity. Data acquisition Data was collected using a free-surface water tunnel with test section dimensions of 0.

R. developed and ran models E. p. 2
Where the paper says it

[9] proposed another FNO variant for predicting isotropic turbulence, finding longterm predictions to be stable. [10] developed a hybrid neural operator network to predict unsteady fluid flows, outperforming vanilla FNO by integrating convolutional long short-term memory and Fourier layers into a single model.

P. and P.R. analyzed data and generated visualizations P. p. 8
Where the paper says it

Given the limitations associated with experimental quantitative flow visualizations, these instantaneous velocity fields do not fully define the state of the turbulent dynamics. The finite resolution, planar observation, and absence of out-of-plane velocity data all prevent the measurements from observing the full state of the dynamics.

R. drafted manuscript All authors reviewed and edited M.G. supervised

Hyperparameters for Model Training

The FCN hyperparameters include an initial learning rate of 10−3, max epochs of 200, batch size of 32, and a total number of parameter 63M The U-Net hyperparameters include an initial learning rate of 10−3, max epochs of 200, batch size of 32, and a total number of parameter 58M The FNO hyperparameters include a Fourier layer width of 80 and Fourier layer modes of 24, with a total number of parameter 59M The DMD model was trained with only 200 of the ten time-step sequences, totaling 2000 samples

Evaluation Speed

Table I shows that all surrogate models tested could produce at least a single time-step forecast faster-than-real-time, potentially enabling future model-based flow control applications The velocity fields are predicted in increments of non-dimensional time approximately equal to t∗ ≈ 0.18 Given a minimum inter-sample time of 5 ms (Re = 2920), all surrogate models tested could produce at least a single time-step forecast faster-than-real-time The mean evaluation times on an NVIDIA A100 are as follows: U-Net 2. p. 1

Where the paper says it

Next, models are trained at a range of Reynolds numbers (Re = 230 to Re = 2920) to investigate the impact of increasingly turbulent and three-dimensional flow phenomena, and the challenges associated with measuring them, on forecast quality. We find that experimentally trained surrogate models can provide meaningful predictions over short time horizons, propagate low-frequency dynamics over longer forecast periods, and achieve faster-than-real-time evaluation.

52 ms, FNO 3. p. 2
Where the paper says it

Particular interest has been shown in predicting fluid flows through operator learning frameworks, such as Fourier neural operators (FNOs) [7]. [8] successfully applied an FNO variant for predicting pressure buildup and gas saturation in multiphase flows.

13 ms, DMD 1. p. 3
Where the paper says it

Using two-dimensional two-component particle image velocimetry (2D2C-PIV) data, common data-driven models are trained to recursively forecast the time evolution of turbulent cylinder wakes. We specifically consider the performance across popular architectures such as fully convolutional neural networks (FCNs), U-Nets, FNOs, and dynamic mode decomposition (DMD) based models.

31 ms, and FCN 2. p. 8
Where the paper says it

U-Nets and FCNs were included to represent common, generalized deep learning frameworks. While U-Nets are often found to surpass FCN performance, FCNs remain commonly deployed for surrogate modeling of fluid flows [26].

02 ms

Experimental Setup Details

Tests were performed in a free-surface water tunnel with test section dimensions of 0.15 m (W) × 0.15 m (H) × 0.61 m (L) An acrylic cylinder of diameter D = 9.53 × 10−3 m was mounted to span the walls of the test section approximately equidistant from the free surface and the tunnel floor Flow speeds were tested ranging from U = 0. p. 5

Where the paper says it

An acrylic cylinder of diameter D = 9. 53 × 10−3 m was mounted to span the walls of the test section approximately equidistant from the free surface and the tunnel floor.

02 m s−1 to U = 0.40 m s−1 The streamwise length of the region of interest (L1) was 0. p. 5
Where the paper says it

Figure 2 depicts the geometry of the recorded region of interest relative to the tunnel boundaries and cylinder. The streamwise length of this region (L1) was 0.

125 m (≈ 13D) with spanwise height (L2) of 0. p. 5
Where the paper says it

The streamwise length of this region (L1) was 0. 125 m (≈ 13D) with spanwise height (L2) of 0.

084 m (≈ 9D) The velocity fields have grid spacing of 9. p. 6
Where the paper says it

fields have grid spacing of 9. 388 × 10−4 m (< 0.

388 × 10−4 m (< 0.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop surrogate models capable of preserving high-frequency spatial details over longer forecast horizons by incorporating skip connections, as seen in U-Net architectures, to mitigate spectral bias that causes predictions to produce overly smooth predictions produced by the FCN model in figure 4.

  2. Implement a method for debiasing Dynamic Mode Decomposition models to better capture physical dynamics by explicitly prioritizing the most energetic modes through Singular Value Decomposition, which can lead to more accurate estimation of planar turbulent kinetic energy.

  3. Design a system that adapts its prediction strategy based on Reynolds number, as the paper shows that the performance is most sensitive to Reynolds number from Re = 230 to Re = 830 for all four models, allowing for tailored forecasting accuracy across different flow regimes.

  4. Integrate physics-informed learning constraints into data-driven losses to enhance model performance by ensuring that the surrogate models adhere better to known fluid dynamics, such as assumptions of divergence-free velocity fields, even when trained on incomplete experimental data.

Abstract

Data-driven modeling shows significant promise for faster-than-real-time forecasting of fluid flows. For real-world engineering applications (e.g., flow control), models must contend with limited, imperfect, and incomplete experimental measurements. In this work, we present an analysis of data-driven surrogate models trained to forecast the time-evolution of experimentally measured cylinder wakes in the subcritical vortex shedding regime. Using a dataset of two-dimensional, two-component particle image velocimetry measurements, we train fully convolutional neural networks, U-Nets, Fourier neural operators, and dynamic mode decomposition-based models to forecast the development of experimentally measured velocity fields. To characterize data-driven approaches contending with transient flow features and limited, imperfect observations, the development of predictions over extended forecast horizons is examined at a fixed Reynolds number (Re = 590). Next, models are trained at a range of Reynolds numbers (Re = 230 to Re = 2920) to investigate the impact of increasingly turbulent and three-dimensional flow phenomena, and the challenges associated with measuring them, on forecast quality. We find that experimentally trained surrogate models can provide meaningful predictions over short time horizons, propagate low-frequency dynamics over longer forecast periods, and achieve faster-than-real-time evaluation. However, the data-driven models struggle to preserve transient flow features and high-frequency energy content when faced with noisy measurements and incomplete state observations. This emphasizes the underlying challenges that remain for data-driven modeling approaches to effectively contend with fluid dynamics in real-world engineering applications, where observations are often imperfect and limited.

Sources

Related papers